Urban scene reconstruction using Geometry-aware Gaussian primitives.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Gaussian splatting advances radiance fields for 3D reconstruction.
• Dynamic urban scenes with rigid and non-rigid objects pose reconstruction challenges.
• Existing methods struggle with geometric integrity in dynamic environments.

Purpose of the Study:

• To develop a geometry-aware framework for dynamic urban scene reconstruction.
• To integrate Gaussian primitives with template meshes for dynamic object representation.
• To ensure geometric integrity and improve appearance generalization in 3D reconstructions.

Main Methods:

• Scene decomposition into a dynamic scene graph.
• Fitting template mesh vertices to observations for topologically consistent models.
• Building Gaussian radiance fields based on template meshes with optimized vertex offsets.
• Projecting appearance attributes to 2D texture space for detailed reconstruction.

Main Results:

• The proposed framework effectively represents dynamic objects using integrated Gaussian primitives and template meshes.
• Achieved superior performance in dynamic urban scene reconstruction compared to mainstream methods.
• Demonstrated accurate reconstruction of geometric surfaces and smoother appearance generalization.

Conclusions:

• The geometry-aware framework provides a foundation for realistic and geometrically complete urban scene reconstruction.
• The integration of Gaussian primitives and template meshes is effective for dynamic elements.
• The method shows significant potential for applications requiring high-fidelity 3D scene understanding.